How accurate is speech recognition in Slovak?

Slovak is recognized better than its close cousin Czech, with a word error rate of 8.45 percent ¹ on Parakeet TDT v3. Whisper large-v3-turbo had 11.51 percent ² on the same 100 read FLEURS sentences. Parakeet leads by about three points and is also faster: 22 times faster than real time ³ against 14 times faster than real time ⁴ for Whisper.

About one word in twelve is wrong with Parakeet. Apple SpeechTranscriber has no row here: in our benchmark it ran only on English, German, French, Spanish, Italian and Portuguese, so for Slovak the choice is between two engines.

Parakeet leads Whisper by three points on Slovak

With punctuation and capital letters counted as a reader sees them, Parakeet's formatted error rate was 10.29 percent ⁵ and Whisper's was 12.58 percent ⁶. The gap narrows to about two points, because Parakeet's text loses more to commas and capitals than Whisper's does. Even so, Parakeet stays ahead.

Published results show the same order

NVIDIA reports 8.82 percent ⁷ for Parakeet TDT 0.6B v3 on FLEURS Slovak, within half a point of our measurement. The Whisper paper lists 11.7 percent ⁸ for Whisper large-v2, almost the same as our 11.51 percent ² for the turbo model. So Whisper has stayed around 11 to 12 percent on Slovak across model versions, and Parakeet is the engine that moved the figure.

Slovak ranks in the upper half of the 25 languages

With Parakeet, Slovak is better than Czech (10.44 percent ⁹), Swedish (9.95 percent ¹⁰ with Whisper) and Romanian (10.00 percent ¹¹), and close to Polish (7.70 percent ¹² with Whisper). It trails Dutch (6.32 percent ¹³) and the large languages such as Russian (4.74 percent ¹⁴). The other Slavic languages in the benchmark vary widely: Slovenian reached only 19.83 percent ¹⁵ with Parakeet.

What to expect when you dictate in Slovak

Expect usable text with a few corrections per paragraph, mostly case endings, proper names and numbers. Parakeet is the clear choice on this language: the lower error and the higher speed point the same way. Yap, which we make, uses Parakeet.

Both engines ran on the device and sent nothing over the network, so the accuracy here does not depend on a connection.

How to read these numbers for Slovak

A word error rate counts every substituted, deleted or inserted word against the written sentence. Slovak has several cases and a word with the wrong ending is a wrong word even when a reader understands it, so the score is stricter than the experience of reading the text. The formatted error rate adds punctuation and capitals, and it is the closest figure to what you see in a finished message. With only 100 recordings per language, a one-point difference between two engines is within what a different sample could change, but a three-point gap like this one is more likely to hold.

What the data does not show

The recordings are read sentences from Wikipedia, spoken in a quiet room, and every engine heard the same 100 of them on one Mac in one run. Real dictation has hesitations, names, numbers and background noise, and none of that is in this benchmark. Numbers and spelling variants are not normalized, so "25" against "twenty-five" counts as an error for every engine alike. Treat the gaps of a point or two between engines as small, and the gaps of five points or more as real.

Where to read more

• All languages in one place: dictation languages, accents and translation and the guide to speech to text in European languages.
• The full rows and every transcript: the on-device ASR benchmark, the published word error rates and the speed data.
• Sibling pages: Czech, Polish, Dutch.